If your agent has a shell, give it a command line

The Model Context Protocol is, in its own words, “an open-source standard for connecting AI applications to external systems” (modelcontextprotocol.io). The project compares it to a USB-C port: one plug for many devices.

The current specification, version 2026-07-28, defines hosts (the AI application), clients (the connectors inside it) and servers (the services that provide data and tools), talking JSON-RPC. A server can offer resources, prompts and tools. Tools are what agents use most: the client asks the server for its list with tools/list, and each tool comes with a name, a description and a JSON Schema for its input.

A command line is not automatically good for agents. It needs a few properties that humans rarely ask for:

• It describes itself as data: --help and a full command catalog as JSON, so the agent never has to guess flags.
• Output is a contract: with --json, exactly one JSON document on stdout, errors as JSON on stderr.
• Exit codes carry meaning: fix the call, ask the human, wait, retry. One number tells the agent what to do next.
• It never waits for input it cannot give: without a terminal, nothing prompts. Destructive steps need an explicit flag such as --yes.
• It can show instead of act: --dry-run prints the exact request without sending it.
• Secrets stay out of the output, also with --verbose and --dry-run, because whatever the agent sees ends up in its context.
• It is easy to install anywhere an agent runs, ideally without a dependency tree.

Anthropic's engineering team described the problem in November 2025: “Most MCP clients load all tool definitions upfront directly into context”, and those descriptions increase response time and cost (Code execution with MCP). Their recommendation is to let the model read tool definitions on demand.

To be fair, clients are catching up. Claude Code now defers MCP tools by default: at session start only tool names and server instructions load, and the full definition comes when Claude searches for the tool (MCP tool search). Other clients may still load everything.

A CLI gets on-demand loading for free, because nothing is in context until the agent runs a command. We measured how much is there to load in tm 0.3.0: the complete catalog (tm schema --json) is 377,203 characters, the overview of all command groups (tm --json) 32,110, the invoices group 104,009. A single command has a median of 934 characters across 157 commands. Method: npx -y @time-momentum/cli@0.3.0 schema --json | wc -c and the same for the other calls, on 10 Oct 2026.

The catalog is not small. That is the point: the agent decides what to load, usually one group or one command.

With MCP, “every intermediate result must pass through the model”, as the same Anthropic article puts it. A list of 500 invoices that the agent only needs to count still lands in its context.

In a shell, the agent filters before it reads. tm payments outstanding --json | jq '[.data[].summary.outstanding_micros] | add / 1000000' returns one number: the outstanding total in euros. tm invoices list --all --json | jq '[.data[] | select(.status == "draft")] | length' counts drafts across all pages. The same lines work in a cron job or a shell script, without any agent.

In MCP, authorization is optional. Remote servers over HTTP should follow the specification's OAuth 2.1 flow; local servers over stdio take their credentials from the environment. For a chat app where every user signs in with their own account, OAuth is the more comfortable option.

A CLI usually works with an API key. That is simple, but it needs care: the key must never show up in what the agent reads. tm keeps it in a local file with mode 600 or an environment variable, exchanges a write key for a token valid for one hour and never prints the key, not even with --verbose or --dry-run. A token from that exchange cannot create new keys or delete the account.

Every call an agent makes through a CLI is a line you can copy, read and run yourself. --dry-run goes one step further and shows method, URL and body without sending anything. That makes the agent's work reviewable before it happens, which matters most for invoices.

MCP has tooling for this too, such as the MCP Inspector for developers. But a tool call is not something you type into your own terminal to see what happens.

A CLI is a package with a version. You can pin it (npx -y @time-momentum/cli@0.3.0), roll back (tm update --to <version>) and know that a script that worked yesterday works today. When the API changes in a way an old version would get wrong, the server rejects that version with a clear error and the update command.

MCP versions the protocol by date and lets servers announce changes to their tool list at runtime. A remote server decides for itself which tools you get today. That is convenient as long as nothing breaks.

Claude Code, Codex, Cursor and most coding agents can run shell commands. A CLI works in all of them without the client having to support anything else, in sandboxes and CI as well.

• Clients without a shell. claude.ai, Claude Desktop and the Claude mobile apps reach external tools through connectors, which are MCP servers (Claude docs). A CLI cannot help there.
• Standard discovery. Every MCP server answers tools/list the same way, and the official MCP Registry lists servers. Every CLI documents itself in its own way.
• Sign-in per user. OAuth lets each person connect their own account without handling a key.
• Non-technical users. Adding a connector in a chat app is easier than installing a command line.

Our users' agents already work in a shell, on client projects. So Time Momentum has no MCP server. It has tm, a command line that runs the whole product: effort, clients, orders, projects, invoices and e-invoices, payments, timesheets and profitability.

tm has no runtime dependencies. Every operation of the REST API that is meant for the command line has exactly one command, 143 operations in version 0.3.0, and a test enforces that with every API change. tm skill returns the usage guide that matches the installed version. In Claude Code, tm skill --install adds a short skill that points to it, and Time Momentum is also available as a Claude Code plugin.

Issuing or cancelling an invoice and deleting need an explicit --yes. Without a terminal, tm stops otherwise.